Distributed multi-camera multi-target association for real-time tracking

Abstract

Tracking and associating different views of the same target across moving cameras is challenging as its appearance, pose and scale may vary greatly. Moreover, with multiple targets a management module is needed for new targets entering and old targets exiting the field of view of each camera. To address these challenges, we propose DMMA, a Distributed Multi-camera Multi-target Association for real-time tracking that employs a target management module coupled with a local data-structure containing the information on the targets. The target management module shares appearance and label information for each known target for inter-camera association. DMMA is designed as a distributed target association that allows a camera to join at any time, does not require cross-camera calibration, and can deal with target appearance and disappearance. The various parts of DMMA are validated using benchmark datasets and evaluation criteria. Moreover, we introduce a new mobile-camera dataset comprising six different scenes with moving cameras and objects, where DMMA achieves 92% MCTA on average. Experimental results show that the proposed tracker achieves a good association accuracy and speed trade-off by working at 32 frames per second (fps) with high definition (HD) videos.